Textual Analysis of Insurance Claims with Large Language Models

This study proposes a comprehensive and general framework designed for exam-ining discrepancies in textual content by large language models (LLMs), broading application scenarios in the �elds of insurtech and risk management and conduct-ing empirical research based on actual needs and real-world data. Our framework incorporates OpenAI’s interface to embed texts and project them into external cat-egories, and utilizes distance metrics to ful�ll discrepancy judgement. To exhibit signi�cant disparities, we design prompts to analyse three relationships: the same information, logical relationship and potential relationship. In our empirical anal-ysis, ChatGPT reveals 22.1% of samples exhibit substantial semantic discrepancy in text statements and require further manual investigation, 38.1% of samples with large differences contain at least one of the identi�ed relationships. The average processing time for each sample does not exceed 4 seconds, and all processes can be adjusted and explained according to actual needs. The backtesting results and comparisons with traditional NLP methods further indicate that our method is both effective and robust.